SPARK - Simple Personal AI Reasoning Kernel

Open Models
3

Hey, sorry — here’s a clearer breakdown of what’s happening.

SPARK = Simple Personal AI Reasoning Kernel.
It’s a small local “brain loop” running on my DGX Spark.
It listens to speech, understands intent with embeddings, and picks which code module (what I call Agent GIFs) should run.

Here’s the pipeline:

  1. Speech In
    Browser captures audio and sends a WAV to my DGX Spark.
  2. Local Speech Recognition
    Whisper Small + Whisper Large run locally on GPU.
  3. Semantic Understanding
    The transcript is embedded using a local embedding model (nomic-embed-text).
    I compare it against embeddings of ~100 modules (“trees”).
  4. Intent Routing (the kernel part)
    SPARK finds the closest match using cosine similarity.
    If confidence is low, it defers to a local 120B model to reason about the safest/correct module.
  5. Module Execution
    The chosen module runs.
    A module is basically an Agent GIF: a self-contained unit (HTML/JS/PHP/text) embedded inside a GIF file.

So in the demo:

  • “What’s the weather in London?” → Weather module
  • “Where is OpenAI based?” → Info module
  • etc.

More detail:

It’s recording speech through the browser, running that through browser speech recognition, saving WAV files, and uploading them to local Whisper Small + Large.
Then it checks which module to execute by comparing embeddings (or asks local 120B to choose).

I’m not using OSS 120B function calling — rolled my own version for now.

In the first example the user asks for weather in London; in the second, where OpenAI is based (California).

Next step is auto-generating modules with GPT-5+ as I have demonstrated in other posts.

download

Modules are basically Agent GIFs.

Would appreciate any feedback — I’m sure I’m making lots of mistakes along the way :slightly_smiling_face: